Enhancing Few-shot Node Classification with High Order Graph Neural Networks

Graph neural networks (GNNs) have achieved significant success in node classification tasks. However, their performance declines when trained on a few examples per class or when applied to unseen classes. Meta-learning tackles this problem by training models across many small learning tasks so that they can quickly adapt to new classes from limited data. In this setting, each task provides a small support set with a few labelled nodes per class, and the model is evaluated on a separate query set of unseen nodes from the same classes. However, applying meta-learning to graphs is particularly challenging due to the interconnected nature of graphs. Existing approaches often enrich tasks with additional contextual information or modify training objectives to better exploit neighbouring nodes and labels through supervised or self-supervised signals. However, the structural complexity of graphs makes it difficult to design stable and transferable tasks, as structural differences across tasks can lead to inconsistent feature representations. We argue that this limitation stems less from the meta-learning framework itself and more from the limited expressive power of standard GNNs. To overcome this, we leverage higher-order GNNs to generate richer node representations during both training and testing, improving the model’s ability to generalize to new classes. Extensive experiments on multiple benchmark datasets demonstrate consistent improvements over state-of-the-art methods. The source code for this project is publicly available at  \url{https://github.com/sirajummprince/HIGH-META}.